cap2vid_with_cnn.py 文件源码

def fully_connected(bottom, n_out, name, reuse=DO_SHARE):
    shape = bottom.get_shape().as_list()
    with tf.variable_scope(name, reuse = reuse):
        # need to flattent he final result, find the dimension that is to be flattened
        dim = 1
        for d in shape[1:]:
            dim *= d
            # print(dim)
        x = tf.reshape(bottom, [-1, dim])

        weights = tf.get_variable('weights', [dim, n_out], tf.float32, xavier_initializer())
        biases = tf.get_variable('bias', [n_out], tf.float32, tf.constant_initializer(0.0))
        logits = tf.nn.bias_add(tf.matmul(x, weights), biases)
        return tf.nn.relu(logits)